---

name: translation-workflows description: "Train, test, and apply CycleGAN, pix2pix, and colorization models with correct dataset modes, checkpoint settings, device choices, and result paths." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Translation workflows

Use this sub-skill for requests to train or test CycleGAN, pix2pix, or colorization models; apply a pretrained generator; resume training; choose CPU/GPU/DDP settings; or diagnose checkpoint/result problems.

Route dataset layout, pair construction, image validation, downloads, and Cityscapes conversion to data-preparation. Route new model/dataset classes and parser/registry changes to customization.

Choose the workflow

• CycleGAN, unpaired domains: --model cycle_gan --dataset_mode unaligned; read references/workflows.md.
• pix2pix, paired domains: --model pix2pix --dataset_mode aligned; read references/workflows.md.
• Colorization: --model colorization --dataset_mode colorization; RGB images become Lab L -> ab pairs internally.
• One-sided generator application: --model test (which selects single) and, when needed, --model_suffix _A or _B.
• Pretrained weights: read references/pretrained-assets.md before any network download.

Operating order

1. Validate the selected data root with data-preparation/scripts/validate_layout.py. The --dataset_mode and --phase determine which folders are read.
2. Choose the model/dataset pair and preserve the model's defaults unless the checkpoint or task requires an explicit override. See references/cli-reference.md.
3. Select the device explicitly through the environment, because the current parser has no --gpu_ids option. Use a CPU-only PyTorch environment or prefix a Linux command with CUDA_VISIBLE_DEVICES= for CPU; prefix with CUDA_VISIBLE_DEVICES=0,1 to choose visible GPUs; use torchrun only for the documented DDP path.
4. Generate or review the command with scripts/build_command.py before running a long job. It prints commands and never starts training or downloads.
5. Keep the training and test architecture settings aligned: --netG, --norm, channel counts, --direction, and dropout/checkpoint suffix choices must match the saved generator.
6. Inspect checkpoints and HTML output rather than treating oscillating GAN losses as a convergence proof. Use references/troubleshooting.md when loading or preprocessing fails.

Outputs and safety

Training writes options, loss logs, checkpoints, and optionally HTML samples under the configured checkpoint root, normally checkpoints/<name>/. Testing writes an HTML result page under results/<name>/<phase>_<epoch>/ (or the configured --results_dir). Confirm those directories before launching a job that may overwrite an experiment name.

The repository imports W&B from its visualizer module even when --use_wandb is false, so the runtime environment must include wandb. W&B logging itself is opt-in; omit --use_wandb when credentials/network access are not available. HTML output is local and can be disabled during training with --no_html.

CUDA/DDP is an optional acceleration path in this skill. The verified inspection environment covers CPU only; do not claim CUDA success from a CPU import. Read the DDP caveat in references/troubleshooting.md before using torchrun.